AI Writing That Sounds Like You: The Voice-Integrity Method

I dictated this into a voice memo after reading a draft my own system had produced, and I stand by every word of it: "It's a simulacrum. It's in the uncanny valley. It's at 80% fidelity and 80% fidelity is just enough to break trust. You see one AI tell creep in there and it just taints the whole work, even if the rest of the work is technically pretty good. So it's one of those things where then if you see one tell you can't trust the whole work. So 80 is pretty darn good, but what I am focused on is, when I say 80, that means it is not 100 fidelity. Your message is diluted to 80% quality, to 80% depth, to 80% nuance. It loses its potency. The real voice carries unique signatures. It's their hallmark."

TL;DR: AI writing that sounds like you starts from your own real words as the spine, treats the tells as a generation constraint rather than a cleanup, and gets checked mechanically before anything ships. I hold this line because 80% fidelity is enough to break a reader's trust.

In writing, 80% fidelity is the exact point where a reader starts to notice something is off. Writers want AI writing that doesn't sound like AI, and most writers never get there, because the standard fix targets the wrong layer.

The Tell Below the Word

The standard fix is vocabulary. The usual advice says delete "delve" and swap out "landscape." That layer is real. Researchers who tracked word frequency across millions of academic papers found that a specific set of AI-favored words spiked sharply after chatbots went mainstream. Words are the visible part of the problem.

The software doing the detecting is a neural network trained on examples. Pangram's technical report describes a model that learned to tell text written by large language models from text written by people, training at scale on human writing paired with machine-written mirrors of the same material. It holds no list of banned words. GPTZero, the detection tool teachers and editors mention by name, began by measuring perplexity and burstiness, meaning how predictable the wording is and how much that varies from sentence to sentence, before moving to a classifier trained on millions of documents in autumn 2023. Both tools now work the same way: trained on labelled examples, they judge the whole pattern of word choice in a passage. The older relative of this software is stylometry, the statistical study of writing style, a field that was unmasking anonymous authors long before chatbots existed.

A large language model, the prediction engine inside every chatbot, writes by choosing the most likely next word over and over. Likely words make text more predictable than a person's, and that predictability is what a trained detector picks up.

A 2025 paper from the Pangram team examined the rewording tools themselves: researchers studied 19 humanizer and paraphrasing tools, software built to reword machine text so it slips past detection. Many existing detectors failed to catch the reworded text. A detector trained with those rewrites in its data still caught them, and it held up when the researchers trained a model to beat it. Rewording a machine draft, by hand or by tool, is a race against the next detector update, and the draft underneath is still machine-written.

Real Words First

The common workflow is generate first, clean up after. You prompt, the machine drafts, and you spend an hour deleting flagged words and swapping in replacements. I ran that loop for months. It's how you end up shipping 80%.

The workflow that holds runs the other way: the draft starts from my real words. I said this to my system: "I also gave you thousands of pre-AI articles that I had written so that you can extrapolate my voice, my tone, how I say things, how I articulate things. Sometimes I think that gets oversimplified into, use somatic language. But you forget I've given you workshops, hours of course transcripts. ... I have given you over a thousand hours of video content to extrapolate my voice."

That corpus is the spine. Articles written before any of this existed, plus over a thousand hours of workshops and course video. A style tip like "use somatic language" skims the surface of that material; the corpus is the material.

When the draft grows out of the corpus, the machine's job shrinks to filling the gaps between things I actually said. When the draft grows out of a prompt, the machine builds the frame and my words become decoration.

This is voice preservation as a practice rather than a hope.

The Mechanical Check

Your own eye is the worst detector of your own tells, because you read what you meant to write. I kept getting tells past my own read, so I built a mechanical check: an automatic pass that runs before anything ships. It blocks rather than advises.

The instructions are deliberately unglamorous: "I want to eliminate all uses of em dashes. Like I have in my AI tells Bible, I want it to be more like that where it's use a colon, a semicolon, or better yet two sentences. I want to look at overuse of the power of three and repetition between paragraphs. I want to look at every sentence and paragraph must earn its spot in the piece."

The em dash earns its own rule. The mark shows up so often in machine text. It has a long, well-documented history worth knowing before you defend it. My fix is the one in the dictation: swap the dash for a colon, or split the thought into two sentences.

The power of three gets the same scrutiny. Two beats build tension; the third beat is where a machine's bias toward tidy symmetry shows through. And the rule that every sentence must earn its spot turns editing from a matter of taste into a pass-fail question the piece either clears or doesn't.

Together these rules form my kill list, and the list is tuned to my voice. Yours will look different, because your tells grow out of your habits. The point is that the check is mechanical. It runs the same way on a tired Friday as on a sharp Monday.

The Stakes and the Proof

I hold the line at 100% for a commercial reason. I said it plainly while wrestling an early version of the system: "Here I am building a system that's supposed to sound like me. You're making things sound like robots. I have extensive documentation on this. We need to fix the system so that every channel has the same level of rigor to the anti-AI tells. ... How can I charge a premium price when the content that comes out of this sounds like every other AI slop?"

The web is filling with AI-generated content. Premium pricing depends on a specific person's voice, the one thing the flood can't mass-produce.

Does the method hold up? "I had specced out a full agent system. It allowed me to finally get four distinct book concepts out of my head and into a place where I can work on them and feel good about them. I feel proud about the work. It sounds like me and it doesn't feel like an AI generated work because so much of my vision, my preferences, my heart is inside of it."

Four book concepts, drafted from real words, at a fidelity I'm proud to sign.

One Pattern to Hunt

Start small. Take your last three posts and hunt one pattern: the power of three or the em dash. Then look at where your drafts begin. If they begin with a prompt, you're decorating a machine-built frame. If they begin with your own words, you have something worth protecting.

When a reader finishes your next piece, whose voice will she actually hear?

Frequently Asked Questions

Why does swapping flagged words fail the mechanical check?

Flagged words sit on the surface, while the check looks at rhythm, repetition, paragraph order, and whether each sentence earns its place. A draft can avoid every risky term and still carry machine-shaped cadence.

What does "corpus as spine" mean in practice?

The draft begins from material you actually made: older essays, transcripts, workshops, memos, and recorded teaching. The system fills gaps between those records instead of inventing a frame and asking your phrases to decorate it later.

Are the em dash and power-of-three rules universal?

No. They are examples from one kill list. The durable rule is that each writer needs explicit checks tied to personal habits, enforced the same way when energy is low.

Why insist on 100% when readers may only sense 80%?

Because trust breaks before a reader can explain the cause. One tell can stain otherwise solid work, and premium work depends on the reader feeling a specific mind behind the page.

Where should a beginner start this week?

Choose one recent post, mark one repeated pattern, and trace the draft back to its origin. If the origin was a prompt, rebuild the opening from something you said before the machine entered the room.